FACTA UNIVERSITATIS (NIS)ˇ SER.: ELEC. ENERG. vol. 19, no. 3, December 2006, 405-428 An Overview of On-Chip Buses Milica Mitic´ and Mile Stojcevˇ Abstract: The electronics industry has entered the era of multi-million-gate chips, and thereXs no turning back. This technology promises new levels of integration on a single chip, called the System-on-a-Chip (SoC) design, but also presents significant challenges to the chip designer. Processing cores on a single chip, may number well into the high tens within the next decade, given the current rate of advancements, [1]. Interconnection networks in such an environment are, therefore, becoming more and more important [2]. Currently, on-chip interconnection networks are mostly imple- mented using buses. For SoC applications, design reuse becomes easier if standard internal connection buses are used for interconnecting components of the design. De- sign teams developing modules intended for future reuse can design interfaces for the standard bus around their particular modules. This allows future designers to slot the reuse module into their new design simply, which is also based around the same stan- dard bus [3]. In this paper we give an overviewof the more popular on-chip bus-based interconnection networks such as AMBA, Avalon, CoreConnect, STBus, Wishbone, etc. The main characteristics of the considered buses in respect to topology, arbitration method, bus-width, and types of data transfers are discussed. Keywords: On-chip interconnection network, on-chip bus, on-chip communication protocol. 1 Introduction Shrinking process technologies and increasing design sizes have led to highly com- plex billion-transistor integrated circuits (ICs). As a consequence, manufacturers are integrating increasing numbers of components on a chip. A heterogeneous SoC might include one or more programmable components such as general purpose processors cores, digital signal processor cores, or application-specific intellectual property (IP) cores, as well as an analog front end, on-chip memory, I/O devices, Manuscript received The authors are with Faculty of Electronic Engineering, University of Niˇs, Aleksandra Medvedeva 14, 18000 Niˇs, Serbia (e-mails: [milicam,stojcev]@elfak.ni.ac.yu). 405 406 M. Miti´cand M. Stojˇcev: and other application specific circuits. In other words, a SoC is an IC that imple- ments most or all the functions of a complete electronic system [4]. On-chip bus organized communication architecture (CA) is among the top chal- lenges in CMOS SoC technology due to rapidly increasing operation frequencies and growing chip size. In general, the performance of the SoC design heavily depends upon the efficiency of its bus structure. The balance of computation and communication in any application or task is, of course, known as a fundamental de- terminant of delivered performance. Usually, IP cores, as constituents of SoCs, are designed with many different interfaces and communication protocols. Integrating such cores in a SoC often requires insertion of suboptimal glue logic. Standards of on-chip bus structures were developed to avoid this problem. Currently there are a few publicly available bus architectures from leading manufacturers, such as Core- Connect from IBM [5], AMBA from ARM [6], SiliconBackplane from Sonics [7], and others. These bus architectures are usually tied to processor architecture, such as the PowerPC or the ARM processor. Manufacturers provide cores optimized to work with these bus architectures, thus requiring minimal extra interface logic. This paper gives an overview of the more popular on-chip standardized buses architectures such as AMBA, CoreConnect, Wishbone, STBus, and others, both from an industrial and research viewpoint. The crucial features, including bus topologies, arbitration methods, bus-widths, and types of data transfers are con- sidered. The rest of this paper is organized as follows: Section II presents background material on CAs, including a survey of typical topologies and communication pro- tocols in use today. Section III, as a central part of this paper, gives an overview of several more popular SoC CAs. In Section IV, for comparison purposes, some common features of the analyzed buses are presented. Concluding remarks are given in Section V. 2 On-Chip Communication Architectures 2.1 Background The design of on-chip CAs addresses the following three issues [8]: 1. Definition of CA topology - defines the physical structure of the CA. Numer- ous topologies exist, ranging from single shared bus to more complex archi- tectures such as bus hierarchies, token ring, crossbar, or custom networks. 2. Selection and configuration of the communication protocols - for each chan- nel/bus in the CA, communication protocols specify the exact manner in which communication transaction occurs. These protocols include arbitra- tion mechanisms (e.g. round robin access, priority-based selection [5, 6], An Overview of On-Chip Buses 407 time division multiplexed access [7], which are implemented in centralized or distributed bus arbiters. 3. Communication mapping - refers to the process of associating abstract system- level communications with physical communication paths in the CA topol- ogy [8]. 2.2 Topologies In respect to topology on-chip communication architectures can be classified as: Shared bus: The system bus is the simplest example of a shared communica- tion architecture topology and is commonly found in many commercial SoCs [9]. Several masters and slaves can be connected to a shared bus. A block, bus arbiter, periodically examines accumulated requests from the multiple master interfaces and grants access to a master using arbitration mechanisms specified by the bus protocol. Increased load on a global bus lines limits the bus bandwidth. The ad- vantages of shared-bus architecture include simple topology, extensibility, low area cost, easy to build, efficient to implement. The disadvantages of shared bus archi- tecture are larger load per data bus line, longer delay for data transfer, larger energy consumption, and lower bandwidth. Fortunately, the above disadvantages with the exception of the lower bandwidth, may be overcome by using a low-voltage swing signaling technique. Hierarchical bus: this architecture consists of several shared busses intercon- nected by bridges to form a hierarchy. SoC components are placed at the appro- priate level in the hierarchy according to the performance level they require. Low- performance SoC components are placed on lower performance buses, which are bridged to the higher performance buses so as not to burden the higher perfor- mance SoC components. Commercial examples of such architectures include the AMBA bus [6], CoreConnect [5], etc. Transactions across the bridge involve ad- ditional overhead, and during the transfer both buses remain inaccessible to other SoC components. Hierarchical buses offer large throughput improvements over the shared busses due to: (1) decreased load per bus; (2) the potential for transactions to proceed in parallel on different buses; and multiple ward communications can be preceded across the bridge in a pipelined manner [8]. Ring: in numerous applications, ring based applications are widely used, such as network processors, ATM switches [5,8]. In a ring, each node component (mas- ter/slave) communicates using a ring interface, are usually implemented by a token- pass protocol. 408 M. Miti´cand M. Stojˇcev: 2.3 On-chip communication protocols Communication protocols deal with different types of resource management algo- rithms used for determining access right to shared communication channels. From this point of view, in the rest of this section, we will give a brief comment related to the main feature of the existing communication protocols. Static-priority: employs an arbitration technique. This protocol is used in shared-bus communication architectures. A centralized arbiter examines accumu- lated requests from each master and grants access to the requesting master that is of the highest priority. Transactions may be of non-preemptive or preemptive type. AMBA, CoreConnect... use this protocol [5, 6]. Time Division Multiple Access (TDMA): the arbitration mechanism is based on a timing wheel with each slot statically reserved for unique master. Special techniques are used to alleviate the problem of wasted slots. Sonics uses this pro- tocol [7]. Lottery: a centralized lottery manager accumulates request for ownership of shared communication resources from one ore more masters, each of which has, statically or dynamically, assigned a number of Xlottery ticketsX [10]. Token passing: this protocol is used in ring based architectures. A special data word, called token, circulates on the ring. An interface that receives a token is allowed to initiate a transaction. When the transaction completes, the interface releases the token and sends it to the neighboring interface. Code Division Multiple Access (CDMA): this protocol has been proposed for sharing on-chip communication channel. In a sharing medium, it provides better resilience to noise/interference and has an ability to support simultaneously trans- fer of data streams. But this protocol requires implementation of complex special direct sequence spread spectrum coding schemes, and energy/battery inefficient systems such as pseudorandom code generators, modulation and demodulation cir- cuits at the component bus interfaces, and differential signaling [11]. 2.4 Other interconnect issues We will point now to several interconnect issues